Traffic Lights Detection And Recognition Based On Deep Learning

Traffic lights are one of the important facilities to ensure the safety of traffic roads and can effectively guide the correct passage of vehicles.Detection and recognition technology can help drivers to correctly interpret traffic lights,reduce the occurrence of traffic accidents,and provide security for autonomous driving.The traditional image processing method has high requirements on the quality of the image,and the machine learning method needs to extract features manually.Deep learning target detection algorithm is difficult to guarantee real-time performance.Aiming at the deficiencies of the above methods,we propose a method that uses the image processing method to detect the candidate regions of traffic lights,and then uses the deep learning algorithm to identify the candidate regions.First,Preprocessing the original image,the ROI is set to reduce the data processing range,and some background information in the image is filtered using morphology,reducing redundant information.In order to improve the color segmentation accuracy,we carry out preliminary segmentation in the RGB space to exclude some color regions,and then performs secondary segmentation of non-zero pixels in the HSV space.In the recognition of traffic signal types,we use a convolutional neural network which commonly used for deep learning.By designing a reasonable convolutional neural network model,the recognition accuracy and real-time performance of the algorithm can be effectively improved.Finally,the performance of this method is verified in the Bosch traffic lights dataset and the Lara traffic lights dataset.The experimental results show that the recognition accuracy and recall rate of this method are both high and meet the requirements of real-time,which has great application value for traffic lights detection and recognition.